Triangulation A field guide to serious AI research
The protocol ← Home

No special software. Three browser tabs and a discipline.

A session runs across the same ordinary interfaces you already use. There is no dashboard, no integration layer, nothing to configure. What makes it a triangulation session rather than a series of isolated queries is the order of operations — and the parts of that order people skip are exactly the parts doing the work.

THE BRIEF Written once CLAUDE Depth CHATGPT Breadth GROK Friction THE CONDUCTOR Reads all three, then decides IDENTICAL · PARALLEL INDEPENDENT NO PLATFORM SEES ANOTHER'S ANSWER BEFORE IT WRITES ITS OWN
The whole method in one picture. Break the independence — by pasting one platform's answer into another, or by rephrasing the brief as you go — and you are no longer triangulating. You are conducting separate interviews and calling the result consensus.

  1. Frame the question

    The session begins before you open a single platform. Write the question out in full — not the topic, the question. Then decide honestly whether this warrants a full session at all: a financial decision, a medical question, a legal clause, a significant creative choice, a plan with real consequences. Not every question deserves the ensemble.

    Framing is not a formality. It is the first act of conducting, and the one that determines the ceiling on everything after it.

  2. Check the brief for loading

    Reread what you wrote and ask whether the answer you're hoping for is visible in the question. If it is, every platform will find it, and you will mistake three confirmations for corroboration. This check takes fifteen seconds and prevents the most common wasted session.

    See Chapter 04 for what neutral phrasing actually looks like.

  3. Brief all three — identically, in parallel

    Write the prompt once. Copy it. Paste it into each platform in its own window, before any of them has responded. Do not rephrase. Do not add context to one that you didn't give the others. If you're working from a document, upload the same document to each — upload, never summarize on their behalf.

    Identical inputs. Independent outputs. That is the briefing. Any variation in what comes back is signal, not noise — but only if the inputs were genuinely identical.

  4. Read everything before evaluating anything

    This is the discipline people skip most often, and skipping it quietly destroys the session. If you evaluate the first response before reading the other two, you have anchored — and the remaining answers get graded against it instead of standing on their own. Read all three. Then start thinking.

    Resist reaching for the one that confirms what you already believed. That one is the most dangerous response in the set.

  5. Map the pattern

    Three questions, in this order:

    Where do they agree? — provisional, not proof.
    Where do they diverge? — the map of genuine uncertainty.
    What did at least one raise that I hadn't considered? — usually the most valuable output of the session.

    Do this at claim level, not document level. Split each response into atomic claims, lay them into a ledger, and mark every row. Three answers you'd describe as "broadly agreeing" will routinely contain two claims that flatly conflict — and those two are why you ran three platforms.

    Flag the outliers explicitly. A lone dissenting answer is either the weakest response in the set or the only one that noticed something. Those are very different, and telling them apart is your job.

  6. Verify what will be acted on

    Before synthesis closes — not after — route the specific factual claims through a grounded source. Numbers, named references, current-status claims, anything that would be harmful or embarrassing if wrong. Unanimity among the three does not exempt a claim from this step. It is precisely the claims all three agreed on that no one is checking.

    Full discipline in Chapter 07.

  7. Second pass — or don't

    A second pass is not asking the same question again. It is a targeted inquiry shaped by what the first round revealed: a specific divergence to resolve, a detail to check, an assumption to test. Reconvene when the question changed, new material arrived, or a significant split needs settling.

    Most sessions do not need one. Running a second pass out of habit is how a session becomes avoidance.

  8. Synthesize — yourself

    Synthesis is not a step the ensemble performs for you, and handing all three answers to a fourth model to "merge" is not synthesis either; it is delegation with extra steps. You take the agreements, the divergences, the raised questions and the verified facts, and you form a view.

    The council advises. You decide. Every platform can reason about the decision. None of them will live with the outcome. That asymmetry of consequence is why the final call is always yours.

  9. Close it — and log it

    Three signals that a session is done: convergence (the genuine divergence has been explored), diminishing returns (each pass produces smaller refinements), or the decision is already made and you're asking for permission. When two appear together, stop.

    A session being done is not the same as a question being answered. Close without deciding when the information doesn't exist yet, a specialist is required first, or time is the next ingredient.

    Typical session · 15–45 minutes

Parallel or serial?

Parallel briefing — everyone at once, independently — is the default, and the only mode that actually triangulates. Use it for open questions, high stakes, and anything that might have dimensions you haven't seen.

Serial briefing moves through the platforms in sequence, using what you learned to shape the next question. It is a legitimate technique for evolving questions and progressive drafts, but be clear-eyed about what it is: once platform two sees platform one's output, their answers are no longer independent, and agreement between them stops meaning anything. Serial work builds depth. It does not produce a fix.

When something feels off

If a session produces only confirmation, check the prompt first. Ninety percent of the time, a session that produced only confirmation was set up to produce only confirmation. Rewrite it with explicit permission to push back, and run it again.

Field note

A chart can guide the band, but it cannot lead the band.